Construction health indicator using physically-informed 1D-WGAN-GP joint attention LSTM-DenseNet method

Abstract

In data-driven prognosis methods, the accuracy of predicting the remaining useful life (RUL) of mechanical systems is predominantly contingent upon the efficacy of system health indicators (HI), typically amalgamated from statistical features derived from collected signals. Nevertheless, the majority of extant HI are beset by two principal shortcomings: (1) during traditional data denoising processes, degradation information from raw data is prone to loss owing to the lack of incorporation of the true physical properties of the data; and (2) the performance evaluation of constructed HI is imbalanced due to the influence of network structures on single models, often resulting in strong performance in only one or two indicators. To overcome such shortcomings, a mechanical health indicator construction method based on physical properties was proposed, termed 1D-WGAN-GP Joint attention LSTM-DenseNet. Firstly, artificial sample data is generated by analyzing the physical properties of the original dataset, which is then used to train the 1D-WGAN-GP model to achieve data denoising. Subsequently, the fusion of the attention LSTM (A-LSTM) network and DenseNet network is utilized to extract crucial feature vectors of HI under varying health conditions from the denoised data. Finally, the extracted feature vectors are used to construct system HI using the Euclidean distance method, and these indicators are used for predicting the system’s RUL. The results indicate that the proposed method outperformed traditional methods in terms of denoising effectiveness. Further, through ablation experiment analysis, the HI constructed by the proposed method demonstrated obvious complementarity in terms of monotonicity, correlation, robustness, and comprehensive evaluation. In RUL prediction applications, the proposed method also exhibited good performance, thereby validating its effectiveness.